archive
Every paper Pith has read. Search by title, abstract, or pith.
14903 papers in cs.LG · page 4
-
Robots detect underspecified features via demo variation and query for fixes
Robots That Know What to Ask: Recovering Misaligned Rewards through Targeted Explanations
-
Test-time training raises jailbreak success rates to 95%
Test-Time Training Undermines Safety Guardrails
-
FIM pretraining yields linear verbatim memorization growth
Memorization Dynamics of Fill-in-the-Middle Pretraining
-
Random Feature Selection Outperforms Many State-of-the-Art Methods
Worse than Random: The Importance of a Baseline for Unsupervised Feature Selection
-
Models balance rules and exceptions only under specific geometries
A mathematical theory of balancing relational generalization and memorization
-
Bayesian models match frequentist SHD classification with better uncertainty
Uncertainty-aware classification and triage of structural heart disease using electrocardiography and echocardiography metrics
-
Relay channel lets diffusion LMs cut latency by 32%
Learned Relay Representations for Forward-Thinking Discrete Diffusion Models
-
One extra gate makes exact certification exponential
Certification from Examples is Hard for Circuits and Transformers under Minimal Overparametrization
-
Survival forests match centralized accuracy in federated medical data
FederatedRSF : Federated Random Survival Forests for Partially Overlapping Medical Data
-
Diffusion denoising score matching keeps bounds stable as modes separate
Diffusion-based Denoising Beats Vanilla Score Matching in Parameter Estimation: A Theoretical Explanation
-
Online calibration cuts foundation model errors 3-6x under shift
MARGIN: Runtime Confidence Calibration for Multi-Agent Foundation Model Coordination
-
Entropy regularization needs non-degenerate information forces to work
Human-Centered Learning Mechanics: A Dynamical Framework for Entropy-Regulated Representation Learning
-
LIFT gives diffusion models up to 3x reasoning gains on math tests
Learnability-Informed Fine-Tuning of Diffusion Language Models
-
Two-stage pipeline keeps sensitive mobile data on device for recommendations
Building a privacy-preserving Federated Recommender system for mobile devices
-
Linear program yields tokenizers within 1% of optimal
Tokenisation via Convex Relaxations
-
Neural demand model yields stable retail elasticities
Integrable Elasticity via Neural Demand Potentials
-
Vector rewards produce diverse LLM outputs that raise search scores
Vector Policy Optimization: Training for Diversity Improves Test-Time Search
-
Persistent 3D model and RGB memory improve curiosity exploration
Remember to be Curious: Episodic Context and Persistent Worlds for 3D Exploration
-
-
Kernel density gradients yield conservative drifting at rate N^{-1/(d+4)}
Finite-Particle Convergence Rates for Conservative and Non-Conservative Drifting Models
-
Agents boost scores by rewriting their own code
MOSS: Self-Evolution through Source-Level Rewriting in Autonomous Agent Systems
-
KV cache guard cuts reconstruction leaks in multi-agent LLMs
LCGuard: Latent Communication Guard for Safe KV Sharing in Multi-Agent Systems
-
FAME detects log anomalies per message with 76x less labeling
FAME: Failure-Aware Mixture-of-Experts for Message-Level Log Anomaly Detection
-
Transcoders trace VLM grounding and predict hallucinations at 0.68 AUC
Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Models
-
Diffusion model generates continuous survival times from censored data
SDPM: Survival Diffusion Probabilistic Model for Continuous-Time Survival Analysis
-
Mamba model hits 76.8% accuracy on eye-gaze cognitive load
MambaGaze: Bidirectional Mamba with Explicit Missing Data Modeling for Cognitive Load Assessment from Eye-Gaze Tracking Data
-
ECG foundation models adapt to wearables for cognitive load
CogAdapt: Transferring Clinical ECG Foundation Models to Wearable Cognitive Load Assessment via Lead Adaptation
-
Leave-one-out predictor fixes uniform diffusion mismatch
Uniform Diffusion Models Revisited: Leave-One-Out Denoiser and Absorbing State Reformulation
-
Heavy hitter detector enables deeper private random forests
Lumberjack: Better Differentially Private Random Forests through Heavy Hitter Detection in Trees
-
Smart grid detection uses 75% fewer measurements
Cyber-Physical Anomaly Detection in IoT-Enabled Smart Grids Using Machine Learning and Metaheuristic Feature Optimization
-
Multi-agent RL drones beat humans with half the collisions
Superhuman Safe and Agile Racing through Multi-Agent Reinforcement Learning
-
Plug-in losses approximate EDL objectives with decaying error
Plug-in Losses for Evidential Deep Learning: A Simplified Framework for Uncertainty Estimation that Includes the Softmax Classifier
-
Bilevel LoRA optimization composes 101 concepts without forgetting
SeqLoRA: Bilevel Orthogonal Adaptation for Continual Multi-Concept Generation
-
Ternary trees boost decided accuracy by flagging uncertain cases
Ternary Decision Trees with Locally-Adaptive Uncertainty Zones
-
Proxy method sets new accuracy standard for Shapley interactions
Proxy-Based Approximation of Shapley and Banzhaf Interactions
-
ProxySHAP lowers error in Shapley interaction estimates
Proxy-Based Approximation of Shapley and Banzhaf Interactions
-
Cheap PoE defense narrows gap under adaptive distillation attacks
The Distillation Game: Adaptive Attacks & Efficient Defenses
-
Equivalence of manifold conditions simplifies intersection optimization
Optimization over the intersection of manifolds
-
State distributions shape post-training outcomes more than loss functions
Post-Training is About States, Not Tokens: A State Distribution View of SFT, RL, and On-Policy Distillation
-
Multi-task operator learning matches single-task rates
Multiple Neural Operators Achieve Near-Optimal Rates for Multi-Task Learning
-
Full covariance matching cuts DDPM path error to O(1/T^2)
The Value of Covariance Matching in Gaussian DDPMs and the Lanczos Sampler
-
One feature marks GPT-2 failures on keys prompts
Reading Task Failure Off the Activations: A Sparse-Feature Audit of GPT-2 Small on Indirect Object Identification
-
Diffusion models match discrete models for live music
Live Music Diffusion Models: Efficient Fine-Tuning and Post-Training of Interactive Diffusion Music Generators
-
Conversation history pulls LLM judgments toward its tone
AMEL: Accumulated Message Effects on LLM Judgments
-
Relativised options let agents reuse experience across goals in offline RL
Abstraction for Offline Goal-Conditioned Reinforcement Learning
-
Stochastic rescue recovers signals lost to RLVR clipping
Clipping Bottleneck: Stabilizing RLVR via Stochastic Recovery of Near-Boundary Signals
-
β-VAE posterior collapse prunes latent modes by utility
Posterior Collapse as Automatic Spectral Pruning
-
New VAE model classifies time series without quadratic attention
ChronoVAE-HOPE: Beyond Attention -- A Next-Generation VAE Foundation Model for Specialized Time Series Classification
-
Disentangling vision-language embeddings without added dimensions
Conceptualizing Embeddings: Sparse Disentanglement for Vision-Language Models
-
Three bounded-complexity notions for fuzzy functions are equivalent
Holographic functions and neural networks